A Fine-Classification Method and its Hardware Acceleration Architecture for Rotation Invariant Multi-View Face Detection

Aiming at detecting faces with all -90°, 90°degree rotation-out-of-plane and 360° rotation-in-plane pose changes efficiently and accurately, this work proposed a fine-classification method and an FPGA-based reconfigurable architecture for rotation invariant multi-view face detection.A coarse-to-fine tree-structured detector hierarchy composed of multiple detector nodes was designed.The proposed method deals with the multi-dimensional binary classification problems in a unified framework by means of a shared output space of multi-components vector.And a novel two-stage boosting method was proposed for training detector nodes.With the exploitation of both the spatial and temporal parallelism of the detection method, a highly parallel reconfigurable architecture template was designed.The reconfiguration of the architecture was evaluated for finding an appropriate tradeoff among the hardware implementation cost, the detection accuracy and speed.Experimental results on FPGA show that high accuracy and marvelous speed are achieved compared with previous related works.A speedup factor ranging from 14.68 to 20.86 for images of size of 160×120 to 800×600 is obtained compared with the conventional software solution on PC.